Object Detection
Finding where objects are in an image and what they are — drawing a labelled box around each one.
Reviewed July 10, 2026Stable
When not to use it
- When one label for the whole image is enough. Detection costs more to label, train, run, and evaluate. Don't buy it if you don't need boxes.
- When you need exact shape. A box around a curved or overlapping object is a crude approximation — segmentation is the right tool.
- On tiny, dense, or heavily overlapping objects without a model specifically chosen for it. Generic detectors degrade badly there.
Reach for something else instead
- Classification for single-subject images.
- Segmentation when boundaries matter more than boxes.
- Classical computer vision — thresholding, template matching, edge detection — is still unbeaten for controlled environments like a factory line with fixed lighting.
Read more on the blog
- The Tempe crash: it saw her for 5.6 secondsThe NTSB found the system detected the pedestrian 5.6 seconds before impact, reclassified her repeatedly, and could not label a person outside a crosswalk. Every safeguard had been disabled for ride smoothness.
- What is computer vision? How machines learn to seeYou unlock your phone with your face, a car reads the road, a scan flags a tumor: all computer vision, the branch of AI that gives machines sight. But capturing pixels is the easy part. Turning a grid of raw numbers into an understanding of what is actually in the picture is the hard problem, and this guide explains how machines learned to do it.
Object detection draws a box around each thing and labels it.
Where classification says only what an image contains, detection says what and where — a labelled bounding box around every object, even many at once. The hard part is doing this in one fast pass rather than scanning every possible location, which is what modern single-shot detectors achieve.
Further reading
- Girshick et al. (2013), Rich feature hierarchies (R-CNN) and Ren et al. (2015), Faster R-CNN — the two-stage lineage.
- Redmon et al. (2015), You Only Look Once — YOLO, and the real-time trade-off.
- Lin et al. (2014), Microsoft COCO — the dataset whose mAP metric everyone quotes and few define.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reporting mAP without saying at what IoU threshold, which makes the number meaningless to anyone else.
- Ignoring non-maximum suppression settings, then wondering about duplicate boxes.
- Labelling inconsistently. Two annotators who disagree about where the box ends will cap your model's accuracy below their agreement rate.
At a glance
FieldComputer Vision
Core idealocate and label each object
Key trade-offspeed vs. accuracy
Metricmean Average Precision (mAP)
DifficultyIntermediate
Where this sits
10 concepts come first. Understanding it opens up 4 more.
7Levelsteps in
10Needs firstconcepts
4Opens up1% of 310
1Areastays here
Learn these firstImage Classification
Computed from the prerequisite graph, not assigned. How this works